Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS5428793 A
Publication typeGrant
Application numberUS 07/435,914
Publication dateJun 27, 1995
Filing dateNov 13, 1989
Priority dateNov 13, 1989
Fee statusPaid
Also published asDE69030425D1, DE69030425T2, EP0428084A2, EP0428084A3, EP0428084B1, US5555417
Publication number07435914, 435914, US 5428793 A, US 5428793A, US-A-5428793, US5428793 A, US5428793A
InventorsDaryl Odnert, Vatsa Santhanam
Original AssigneeHewlett-Packard Company
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Method and apparatus for compiling computer programs with interproceduural register allocation
US 5428793 A
Abstract
Optimization techniques are implemented by means of a program analyzer used in connection with a program compiler to optimize usage of limited register resources in a computer processor. The first optimization technique, called interprocedural global variable promotion allows the global variables of a program to be accessed in common registers across a plurality of procedures. Moreover, a single common register can be used for different global variables in distinct regions of a program call graph. This is realized by identifying subgraphs, of the program call graph, called webs, where the variable is used. The second optimization technique, called spill code motion, involves the identification of regions of the call graph, called clusters, that facilitate the movement of spill instructions to procedures which are executed relatively less often. This decreases the overhead of register saves and restores which must be executed for procedure calls.
Images(7)
Previous page
Next page
Claims(4)
What is claimed is:
1. A method of operating a general purpose data processor having a plurality of machine registers, a sub-set thereof being assigned for use as interprocedural registers, so as to allow more efficient allocation of said interprocedural registers when said data processor is executing a computer program comprising a plurality of procedures, at least one of said procedures operating on a global variable, said method comprising the steps of:
building a program call graph, said program call graph comprising a set of nodes, each said node representing a procedure, interconnected by directional edges to other said nodes, each said edge representing a call from a first procedure to a second procedure, the node representing said first procedure being the ancestor of the node representing said second procedure and the node representing said second node being the descendent of the node representing said first procedure;
defining webs corresponding to global variables, each said web corresponding to a global variable, each said web comprising a collection of program call graph nodes such that said corresponding global variable is accessed in at least one node in said web and such that, for each node in said web, said corresponding global variable is not accessed in any ancestor node not in said web, and said global variable is not accessed by any descendant node not in said web;
determining an order for said webs; and
assigning said global variables to interprocedural machine registers according to the order of said webs corresponding to said global variables in said determined order.
2. The method of claim 1 wherein said priority order assigned to said webs is determined by the frequency of use of the global variable corresponding to each said web.
3. The method of claim 1 wherein said priority order assigned to said webs is determined from profile information collected by running said program with exemplary input data on a data processing system.
4. The method of claim 1 wherein said step of assigning interprocedural machine registers comprises assigning different machine registers to webs having common program call graph nodes.
Description
COPYRIGHT NOTICE

A portion of the disclosure of this patent document contains material to which a claim of copyright protection is made. The owner has no objection to the facsimile reproduction of the patent document or patent disclosure as it appears in the Patent and Trademark Office patent file or records, but reserves all other rights whatsoever.

BACKGROUND OF THE INVENTION

This invention relates to computer program compilers and more particularly a computer program compilation system that supports register allocation across procedure and compilation unit boundaries where there are global variables and a limited number of registers for storage and manipulation of data.

In a traditional compiler, register allocation is performed on each procedure one at a time. In some compilers, the register allocator has access to register allocation information from other procedures within the same compilation unit. The compiler can use this information to improve the register allocation in the callers of these routines. This type of technique is limited in scope to the procedures of a single compilation unit.

The traditional intraprocedural register allocation process is effective, but in the absence of interprocedural information the following situations occur:

Local values in different procedures are assigned to the same register. As a result, procedures must execute code to save and restore these registers in order to preserve the values needed by the calling procedure.

Global variables are referenced out of different registers in different procedures. This requires a modified value of a global variable to be stored to memory before any procedure call, and loaded back from memory before any subsequent use of that variable. This also requires each procedure which references that variable to load the variable from memory if it is used before being redefined, and to store the global variable to memory before the exit point if the variable is modified within that procedure.

For most programming languages, improving this situation is complicated by the need to support multiple compilation units. For example, if one wishes to keep a certain global variable in a register when compiling module A, one must ensure that any reference to that variable in a different module uses the same register.

One possible solution is to delay register assignment until link time, when the code for the entire application is visible. This solution is difficult to implement with traditional compiler architectures, however, because of the need for dataflow and live range information at register allocation time. Moreover, computing this information would create an unreasonable delay each time a user needed to re-link an application.

There are two known significant research efforts that have addressed the weaknesses of procedure-at-a-time register allocation. The first was carried out at DEC's Western Research Lab in 1986 and described by David W. Wall in an article entitled "Global Register Allocation At Link Time" in the Proceedings of the SIGPLAN '86 Symposium On Compiler Construction, SIGPLAN Notices, Vol. 21, No. 7, July 1986, pages 264-275. In this technique, the compiler does a simple register allocation on each procedure and generates register relocation information for the linker. The user may optionally enable interprocedural register allocation at link time. To promote a variable to a register, the linker only needs to follow the prescribed relocation actions. This technique showed some good results. Some benchmarks improved by as much as 8% on a 64-register RISC machine, with a majority of the benefit attributed to the promotion of global variables.

Global variable promotion is an optimization technique where memory references to global variables are converted into register references. In effect, the global variable is promoted from being a memory object to a register object. Traditional compilers sometimes promote global variables to registers locally within a procedure. Such locally promoted global variables are still accessed out of memory across procedures. Before procedure calls and at the exit point, the compiler inserts instructions to store the register containing the promoted global variable back to memory. Similarly, just after procedure returns and at the entry point, the optimizer inserts instructions to load the promoted global variable from memory to register.

The second significant research effort was produced at MIPS Computer Systems and described by Fred C. Chow in "Minimizing Register Usage Penalty at Procedure Calls" published in Proceedings of the SIGPLAN '88 Conference on Programming Language Design and Implementation, July 1988, pages 85-94 and also in an article authored with others in "Cross-Module Optimizations: Its Implementation and Benefits" published in the Proceedings of the Summer 1987 USENIX Conference, pages 347-356. In the MIPS system, the multiple compilation unit problem is solved by exposing an intermediate code representation to the user. Then, instead of linking object code, the user must link the intermediate code files into a single, large intermediate program file. The intermediate code linker then completes the code generation and optimization process. As part of this process, the optimizer tries to minimize register spill by performing register allocation on procedures in a reverse hierarchical order and propagating register usage information upwards in the call graph. This technique showed generally positive results, although there were exceptions noted. In one example discussed, this process resulted in object code which executed more slowly than a version compiled without interprocedural register allocation. There are other computer systems and compilers that have been implemented which use a similar technique within a single compilation unit.

On many contemporary computer architectures, machine registers are divided by software conventions into three classes: status registers, caller-saves registers, and callee-saves registers.

Status registers are registers which are designated to hold specific values which may not be used to hold variables or other temporary values. Examples include a stack pointer and a global data pointer.

Caller-saves registers are registers which may be used within a procedure to hold values, but these values are not guaranteed to remain unchanged after executing a call to another procedure. These registers may be used by a procedure without being preserved in memory before they are used. The name "caller-saves" refers to the fact that the caller of a procedure must save any needed values in these registers so the called routine may use those registers.

Callee-saves registers are registers which may be used within a procedure to hold values, and these values are guaranteed to remain unchanged after executing a call to another procedure. However, the values in these registers must be spilled before they are used and then restored to the register before exiting the procedure. The name "callee-saves" refers to the fact that the called routine is responsible for saving these registers before they are used.

In the absence of interprocedural information, callee-saves register spilling is necessary in every procedure which needs to use a register of that class. This creates significant overhead in many programs.

Some other references to related work include:

"LISP on a Reduced Instruction Set Processor: Characterization and Optimization", by P. A. Steenkiste of Stanford University Computer Systems Laboratory, PhD Thesis, Chapter 5, March 1987. This approach is similar to that of MIPS, except that it reverts to ordinary intraprocedural register allocation when the interprocedural registers are exhausted in upper regions of the call graph.

"Data Buffering: Run-Time Versus Compile Time Support" by Hans Mulder, Proceedings of the 3rd International Conference on Architectural Support for Programming Languages and Operating Systems, Apr. 3-6, 1989, pages 144-151. This approach is also similar to that of MIPS except that it is limited in scope to single compilation units.

"The Impact of Interprocedural Analysis and Optimization in the Rn Programming Environment" by Keith D. Cooper, Ken Kennedy, and Linda Torczon of Rice University. Published in the ACM Transactions on Programming Languages and Systems, October 1986, pages 491-523. This paper describes a program compiler which computes interprocedural optimization information, but does does not address the register allocation problem.

Hewlett-Packard's Apollo Division uses an interprocedural register allocation scheme within a single compilation unit in their DN10000 architecture compilers. As with the references above, except for the DEC paper, this approach does not attempt to keep global variables in registers across procedures.

What is needed is a method and apparatus for optimizing register usage where there is a limited number of available register resources in a computer processor and where a plurality of procedures and variables are involved.

SUMMARY OF THE INVENTION

According to the invention, two specific optimization techniques are implemented by means of a program analyzer used in connection with a program compiler to optimize usage of limited register resources in a computer processor. The first optimization technique, called interprocedural global variable promotion allows the global variables of a program to be accessed in common registers across a plurality of procedures. Moreover, a single common register can be used for different global variables in distinct regions of a program call graph. This is realized by identifying subgraphs of the program call graph, called webs, where the variable is used. The second optimization technique, called spill code motion, involves the identification of regions of the call graph, called clusters, that facilitate the movement of spill instructions to procedures which are executed relatively less often. This decreases the overhead of register saves and restores which must be executed for procedure calls.

The program analyzer according to the invention reads summary files produced by a compiler modified to create summary files containing, for each procedure of a source code file, global variable usage information, register need information, and names of called procedures. The compiler is run on each source code file separately to produce separate summary files. The program analyzer computes interprocedural register allocation information from the summary files and writes it out to a program database file.

The program analyzer builds a single program call graph (PCG) from all the summary files. The PCG consists of a set of nodes, each representing a procedure, interconnected by directional edges, each representing a call from a first procedure to a second procedure.

Traditional intraprocedural register allocators have in the past employed data structure known as def-use chains to represent live ranges of variables. Def-use chains are analogous to spider webs linking equivalence classes of definitions and uses of a variable. Consequently, def-use chains are sometimes referred to as "webs". Webs have not been employed in the past in any interprocedural register allocation techniques. In other words, live ranges for global variables have not been computed across procedure boundaries.

In order to facilitate global variable promotion, the program analyzer identifies webs for for selected global variables (the global variables selected are those variables that are eligible for assignment to an interprocedural machine register). A web for a single global variable is a collection of PCG nodes such that the global variable is accessed in at least one node of the web and such that, for each node in the web, the global variable is not accessed in any ancestor node not in the web, and the global variable is not accessed by any descendant node not in the web. Multiple webs may be identified for a single global variable.

The program analyzer then prioritizes the webs according to frequency of use of the corresponding global variable within nodes of the web. The webs can optionally be prioritized based on profile information collected from an earlier run of the compiled source files.

The program analyzer then assigns the first available interprocedural machine register to the selected webs in priority order. In assigning interprocedural machine registers to the selected webs, the program analyzer ensures that webs that have common PCG nodes are assigned different machine registers. The interprocedural registers assigned to the webs are chosen from a limited sub-set of machine registers designated for preserving values across procedure calls (callee-saves registers).

Unlike previous approaches, this method of interprocedural register assignment allows a single register to be used for different purposes in distinct regions of the PCG. Specifically, a single interprocedural register can be used for the promotion of different global variables in different regions of the PCG. This allows a larger number of global variables to be promoted than the approach described by David W. Wall. Moreover, with our method, global variables are not be promoted in the regions of the PCG in which the variable is not used.

Another function of the program analyzer is to facilitate the reduction of the overhead associated with saving and restoring callee-saves registers. This overhead is mitigated through spill code motion.

In order to facilitate spill code motion, the program analyzer first identifies clusters of nodes of the PCG. A cluster is a collection of PCG nodes such that there exists a unique root node of said cluster only through which every other node in the cluster can be called. Profile information collected from an earlier run of the compiled source files may be used to aid cluster identification.

Interprocedural machine registers are assigned to each cluster node, according to the register need for the corresponding procedure and as restricted by the cluster organization.

The root node of each cluster is designated to execute machine instructions to preserve the values of the interprocedural registers assigned to nodes of that cluster. The machine instructions are executed upon calls to the cluster root node so that other nodes within the cluster need not execute the machine instructions.

The assignment of interprocedural machine registers to global variable webs and cluster nodes are finally written out by the program analyzer to the program database.

The summary files read by the program analyzer are produced in conjunction with intermediate files by a first phase of compiler operation. In a second phase, the program database file and intermediate files created by the first phase, are processed to produce individual object code files with the requisite interprocedural register assignments. In this stage, the pseudo-register operands used in the intermediate code (read from the intermediate files) are mapped into machine registers by a register allocator. The register allocator uses the interprocedural machine registers specified in the program database file to map certain pseudo-registers to machine registers.

For each procedure corresponding to a node of a web that was assigned an interprocedural machine register by the program analyzer, all memory references to the corresponding global variable are converted into interprocedural register references. At the root nodes of webs that were assigned an interprocedural machine register, instructions are added at the entry point to load the value of the corresponding global variable from memory into the interprocedural register and at the exit point to store the value back to memory. Additionally, machine instructions (spill code instructions) are added to preserve the value of the interprocedural register across calls to that root node. These transformations effectively promote the storage class of the selected global variables to register i.e. they result in global variable promotion.

For each procedure corresponding to a node of a cluster that was assigned interprocedural machine registers by the program analyzer, certain pseudo-registers are mapped into those interprocedural machine registers instead of the ordinary callee-saves registers. Machine instructions are added at the root nodes of clusters designated to preserve the values of the interprocedural registers assigned to nodes of that cluster. These transformations effectively result in spill code motion.

The invention will be better understood by reference to the following detailed description in connection with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of a computer system incorporating the invention.

FIG. 2 is a block diagram of a computer program compilation system in accordance with the invention.

FIG. 3 is a block diagram of a register allocator used in accordance with the invention.

FIG. 4 is an internal view of the compilation system operational in accordance with the invention.

FIG. 5 illustrates a sample program call graph and the webs identified for three global variables.

FIG. 6 is a table describing the usage of the global variables at different nodes of the program call graph illustrated in FIG. 5.

FIG. 7 illustrates clusters of nodes within the same program call graph shown in FIG. 5.

FIG. 8 is another sample program call graph to illustrate a potential cluster identification problem.

FIG. 9 is another sample program call graph to illustrate a potential web identification problem.

DESCRIPTION OF SPECIFIC EMBODIMENTS

The invention is described by a specific embodiment as might be implemented on a Hewlett-Packard Precision Architecture computer system. The invention may be applied to any computer system where there is a finite number of registers and more particularly to computer systems where the number of registers available for computation and other processing is limited. This invention applies particularly to computer systems that categorize the available registers by software convention into callee-saves and caller-saves register partitions. The invention operates on source files, which are human readable high-level program fragments which comprise procedures.

Referring to FIG. 1, there is shown a computer system 10 incorporating the present invention. A compiler 12, is resident in a computer 14, including a processor, and a program analyzer 16 is resident in the computer 14 in accordance with the invention. The invention operates as follows:

A plurality of source code files 18, 20, 22 are supplied one at a time as input to the computer 14. The computer 14, through the use of the compiler 12 and the program analyzer 16, processes the source code files 18, 20, 22 to produce object files incorporating interprocedural register allocation optimization in accordance with the invention. The object files are linked into an executable file 24 by a linker 17. The executable file 24 may be run on the computer 14 or on another machine of the type for which the compiler has generated the executable file 24.

Referring to FIG. 2, there is shown a block diagram of a portion of an apparatus 26 for performing interprocedural register allocation in accordance with the invention. The apparatus 26 comprises a scanner 28, for sequentially reading source code files 18, 20, and 22, a parser 30, coupled to receive from the scanner 28, lexically checked tokens of the source code files one at a time for verifying syntactic correctness. In addition, a translator 32, coupled to receive the output of the parser 30 is operative to check for semantic correctness and to generate an intermediate code representation 34, 36, 38, for each of the source code files, respectively. The intermediate code is an ordered list of procedures translated from the source code. Each procedure consists of a sequence of machine instructions 108, each of which consists of an opcode 100 and pseudo-registers 102, 104, 106 as shown in FIG. 3. The intermediate code is temporarily stored for subsequent use or regenerated later. Each unit of the intermediate code is analyzed by a variable usage analyzer 40, register need analyzer 42, and a procedure call analyzer 44.

The variable usage analyzer 40, identifies the global variables used in each procedure, and the nature of the access to the global variables.

The procedure call analyzer 44 identifies the names of the called procedures and the nature of the calls to each procedure.

The register need analyzer 42 estimates machine register need by examining the intermediate code.

The summary file generator 46 outputs the results produced by the analyzers to summary files 48, 50, 52 in accordance with the invention. The summary file contains a record for each procedure in the corresponding source files. The summary files 48, 50, 52 are processed serially by a program analyzer 54 which analyzes the summary files as herein below described. The results of the program analyzer analysis is stored in a program database 56.

Optionally, profile information from a profile information file 58 is also provided to the program analyzer 54. The profile information is collected from an earlier run of the result of a previous compilation 60 of the same source files 18, 20, and 22.

Referring to FIG. 3 there is shown an illustration of a register allocator 62 in accordance with the invention. The register allocator 62 is operative to receive the information in the program database 56 and the intermediate code 34, 36, 38 in the form of procedures. Each procedure consists of a sequence of instructions 108, each of which consists of an opcode 100 and pseudo-registers 102, 104, and 106. The register allocator 62 is responsible for converting memory references to promoted global variables into interprocedural register references. The register allocator 62 is also responsible for replacing all of the pseudo-registers 102, 104, 106 in the intermediate code with machine registers, some of which may be interprocedural registers 110, 113. The register allocator is also responsible for adding spill code 116, 118 to preserve the values of certain registers, including the interprocedural registers in root nodes of clusters and webs as herein after explained.

The result of register allocation for each set of procedures is output to an object file 120, which is in turn processed by the linker 17 (FIG. 1) into an executable file 24 (FIG. 1).

Referring to FIG. 4, there is shown an internal overview of the compilation system 122 with source files 18, 20, 22, summary files 48, 50, 52, intermediate files 34, 36, 38, program analyzer 54, profile information 58, program database 56, object files 120, the linker 17, run time libraries 19, an executable file 24. The compiler comprises a first phase 12A and a second phase 12B. The compiler first phase shown in FIG. 4 comprises the scanner 28, the parser 30, and the translator 32 shown in FIG. 2. The compiler second phase 12B shown in FIG. 4 includes the register allocator 62 shown in FIG. 3.

The collection of summary files 48, 50, 52 produced by the compiler first phase 12A is exposed to the program analyzer 54. The program analyzer is responsible for constructing a program call graph 130 (FIG. 5) for the program, identifying webs and clusters for the program call graph and making decisions about how interprocedural registers are to be allocated to the webs and clusters across that graph 130.

The program call graph 130 comprises nodes A, B, C, D, E, F, G, H, and I with directional edges 132, 134, 136, 138, 140, 142, 144, 146, and 148 connecting the nodes. Webs 150, 152, 154, and 156 are identified for the program call graph 130 in FIG. 5 and clusters 158, 160, and 162 are identified on the same program call graph 130 in FIG. 7 as hereinafter explained.

An interprocedural register is allocated to each selected web. A plurality of interprocedural registers are allocated to nodes of each cluster. These allocations are recorded in the form of register actions to a program database 56 of information about the program. The register allocator 62 uses information from the program database 56 to guide register allocation on each procedure in the intermediate file, one at a time.

The program analyzer 54 has been designed with the availability of profile information 58 in mind. In a particular embodiment, the implemented algorithms are not dependent on profile information, but some of the heuristics used in the program analyzer can be improved with profile information 58.

The Program Analyzer

The program analyzer 54 generates the program database 56 by completing the following tasks, in order.

[1] Build a program call graph 130 (FIG. 5).

[2] Identify webs for global variables and select webs for which interprocedural registers are to be allocated.

[3] Assign interprocedural registers to the webs selected in step 2.

[4] Identify clusters in the call graph.

[5] Pre-allocate registers to the nodes of each cluster identified in step 4 and identify the registers that must be spilled at the root of each cluster.

[6] Compute register actions for each node of the program call graph and write these out to the program database.

Global Variable Promotion Strategy

Interprocedural register promotion is an optimization that attempts to maintain certain important user global variables in registers across procedure boundaries.

The preferred embodiment makes use of the compilation system shown in FIG. 4 to automatically promote key global variables to registers across procedure boundaries. The compiler first phase 12A is responsible for communicating the callee-caller relationships and global variable usage for each of the source files 18, 20, 22 to the program analyzer 54 through summary files 48, 50, 52.

The program analyzer 54 collects the global variable usage information across all source files that make up the program. It determines which globals are ineligible for register promotion and of the remaining global variables, selects those that are the most heavily referenced, for register promotion.

The program analyzer 54 partitions the references to a global variable into disjoint webs and consider each web individually for register allocation.

A web, for the purposes of interprocedural global register allocation, identifies a minimal set of nodes in the call graph over which an eligible global variable could reside in an interprocedural callee-saves register (without requiring caller spill). Typically, the nodes of a web would be a connected subgraph which use or define the web variable. The subgraph is selected such that no ancestral node or descendent node in the call graph references the same web variable.

Each web can have one or more entry nodes. These entry nodes are basically the root nodes of the web subgraph. Typically, an entry node would contain references to the global variable while none of its direct predecessors would. All descendant nodes of each entry node that either have a local reference or a child reference to the global, are part of the web. If a web is selected for promotion, at every node of the web, references to the global memory variable would be replaced by a register reference to a reserved interprocedural callee-saves register which would be unavailable for normal intraprocedural register coloring.

Additionally, at the entry nodes of the web, the global variable is loaded into the reserved callee-saves register on entry and stored back on exit (e.g. instructions 115 and 117 in FIG. 3). (Note that this allows a global variable whose value has been initialized at compile time to be safely promoted to a register.) For this approach to work, entry nodes are required to have ONLY external predecessors nodes and internal nodes are not allowed to have ANY external predecessors.

To meet this requirement, special entry nodes may need to be added to the web. For example, consider the program call graph 200 in FIG. 9. Suppose a global variable g3 is referenced in procedures S 202 and U 204, but not in procedure T 206. Suppose further that the web for g3 were to include only nodes S and U. If that web is assigned an interprocedural register, the instructions for procedure U will reference the value of g3 from that interprocedural register. When U is called from T, however, no instructions will have been executed to place the most recent value of g3 in that interprocedural register. This will be an incorrect translation of the source code. To solve this problem, procedure T is added to the web and designated to be an entry node.

The webs for a global variable are identified by using dataflow sets (bit vectors of numeric global identifiers, L-- REF, C-- REF, and P-- REF) at each node of the call graph. These dataflow sets are computed by propagating the local reference information at each node throughout the call graph. After the webs are built, an interference graph is constructed. In an interference graph, each node represents a web and each edge represents an interference between two webs. Two webs interfere with each other if the respective global variables are simultaneously `live` at one or more nodes in the call graph (i.e. if the two webs have common call graph nodes). A global variable is `live` at a node, if there are references to that global variable either at that node or in an ancestral or descendent node in the program call graph.

Note that by definition, two different webs of the same global variable will never interfere with each other. The nodes of the interference graph will be assigned (colored with) one of the reserved callee-saves registers in a manner which avoids assigning the same register to adjacent nodes in the interference graph. Before coloring, the webs are sorted into an order based on heuristics. If a node is colored successfully, then the corresponding web is in effect promoted to a register. Register actions will then be assigned to the nodes comprising a colored web. These register actions are used to direct the register allocator on how to promote the global variable to a register. The register actions for a node could include one or more of the following directives.

1. Make callee-saves register R unavailable for intraprocedural use.

2. Add save and restore code for register R (e.g. instructions 116 and 118 in FIG. 3). Also, add load and store instruction for global variable G (e.g. instructions 115 and 117 in FIG. 3) to and from callee-saves register R on entry and exit (e.g. for variable g1 in procedure B in FIG. 5). Note that a load or store instruction can be generated from the information in the program database.

3. All references to global variable G (e.g. g1 in FIG. 5) are to be replaced with references to callee-saves register R (e.g. r17 in FIG. 5).

The register actions for each node are written out to the program database 56 by the program analyzer 54 and later queried and used by the compiler phase 2 (12B in FIG. 4).

To better understand the algorithm described below, consider FIG. 5 and FIG. 6. The nodes of the program call graph 130, labeled A through I represent procedures and the edges 132, 134, 136, 138, 140, 142, 144, 146, 148 between these nodes represent procedure calls. FIG. 6 describes how three global variables g1, g2, g3 are accessed by the different nodes of the program call graph.

The webs 150, 152, 154, 156 are identified for these three global variables. Note that there are two separate webs 152, 156 for the global variable g2. For this example, all four webs can be promoted using just two callee-saves registers, r17, r18. These callee-saves registers could correspond to the interprocedural registers 110 and 113 shown in FIG. 3.

Different webs for the same variable may be assigned different registers. This is the case for web 4 156 and web 2 152 for global variable g2, which are assigned registers r17 and r18 respectively.

The compiler second phase 12B is responsible for converting memory references to global variables corresponding to promoted webs into register references. For example, memory references to global variable g2 are converted into references of register r17 in procedures C, F, and G. Additionally, at the entry procedure C for web 2 152, code is inserted at the beginning to load the global variable g2 into the register r17. Code is also inserted at the end of procedure C to store back the global variable g2 from register r17 to the memory location in the computer associated with that variable.

Interprocedural Global Register Promotion Algorithm

The following pseudo-code provides guidance for coding the algorithms discussed herein. This pseudo-code is based on the "C" programming language.

Input

A program call graph of `n` nodes. It is assumed that the representation of the call graph allows easy identification of a node's successors and predecessors.

For each node of the call graph, a list of global variables that are explicitly referenced, a set of heuristically assigned reference frequencies for those globals, and the set of global variables (not necessarily a subset of those referenced) whose address have been taken or are otherwise ineligible for promotion. This information is computed by the variable usage analyzer 40 shown in FIG. 2.

Output

For each node of the call graph, intraprocedural register actions required to maintain a selected set of singleton global variables in registers within certain regions of the call graph.

Main Data Structures

L-- REF [i] The dataflow set representing Local-- References. For each node i, L-- REF[i] represents the set of eligible global variables (indices) referenced locally at that node.

P-- REF [i] The dataflow set representing Parent-- References. For each node i, P-- REF[i] represents the set of eligible global variables (indices) referenced by all parent nodes of node i. A parent node is any node in the call graph from which there is a forward call path (an invocation sequence without procedure returns) to node i.

C-- REF [i] The dataflow set representing Child-- References. For each node i, C-- REF[i] represents the set of eligible global variables referenced by all child nodes of node i. A child node is any node in the call graph to which there is a forward call path (an invocation sequence without procedure returns) from node i.

R-- Action [i,g] Register Actions (1 . . . 3) at node i for promoted global variable

Ref-- Count [g,i] The reference count for variable g at node i

Global-- Table [g] For each unique eligible global g, there is an entry in the Global-- Table. Each entry for a global variable in the Global-- Table has a unique index that is its numeric identifier (which is used in the dataflow sets). Each entry also contains the set of call graph nodes which access that global. The organization of this symbol table allows for easy entry, deletion and translation between the global variable name and entry index.

Web-- Table [w] For each web (identified by its entry index in this table), the Web-- Table entry contains the numeric global index, the entry nodes of the web, the other member nodes of the web, and the total reference count of the global over the nodes in the web.

Web-- Interferences [w] This is the web interference graph. The Web-- Interferences entry for each web contains the web numbers of all other webs that it interferes with.

______________________________________Pseudo-codeAllocate-- Global-- Registers( )Initialize-- Global-- Table( );Compute-- P-- REF ( );Compute-- C-- REF ( );for each global-- id in Global-- Table do {Compute-- Webs (global-- id);}Compute-- Web-- Interferences ( );Color-- Webs ( )Compute-- Reg-- Actions ( );}Initialize-- Global-- Table ( ){for each summary file do {Open the summary file for reading;for each procedure p in the summary file do { /* see File Structure Description given later */ for each Global-- Usage-- Info-- Record do {  if (global exists in Global-- Table[ ]) then   i = Global-- Table[ ] index for global;   else   Add Global-- Var-- Name to Global-- Table[i];  Global-- Table[i].Ref-- nodes += {p};  Ref-- Count[i,p] = Weighted-- Reference-- Count; }}Close the summary file;}/*Now the globals ineligible for register promotionneed to be removed from the Global-- Table[ ]. A globalis ineligible if it's address has been taken, if it'sa volatile variable, or if there have been abnormalreferences to it, as indicated by the Global-- Attributes.*/for each entry in Global-- Table[ ] do {if (global ineligible for promotion) then {  Remove Global-- Var-- Name From Global-- Table[ ];}}/* The variable names present in the Global-- Table are/* now the candidate globals that are eligible for/* interprocedural register allocation */for each global i in Global-- Table do {for each proc p in Global-- Table[i].Ref-- nodes do {  L-- REF[p] = L-- REF[p] + { i };}}}Compute-- P-- REF ( ){repeat {Changes := false;for each node n in Depth First Search order doCalc-- P-- REF (n);} until no Changes;}Calc-- P-- REF (i){temp-- P-- REF := .0.;for each predecessor p of i do {temp-- P-- REF := temp-- P-- REF ∪ (P--REF[p] ∪ L-- REF[p]);}if (temp-- P-- REF ≠ P-- REF[i]) then {P-- REF[ i] := temp-- P-- REF;Changes := true;}}Compute-- C-- REF ( ){repeat {Changes := false;for each node n in reverse Breadth First Search order doCalc-- C-- REF(n);} until no Changes;}Calc-- C-- REF (i){temp-- C-- REF := ;for each successor s of i do {temp-- C-- REF := temp-- C-- REF ∪ (C--REF[p] ∪ L-- REF[p]);}if (temp-- C-- REF ≠ C-- REF[i]) then {C-- REF[i] := temp-- C-- REF;Changes := true;}}Compute-- Webs (g){/*Build webs by first identifying primary entry nodes. Arequirement for a primary entry node is that theglobal-- id be part of its L-- REF set but not the P--REFset. Any references to the global found within astrongly connected component of the call graph will getconsidered last.*/Ref-- nodes := Global-- Table[g].Ref-- nodes;Scc-- refs := false;analyze-- refs:for each node p in Ref-- nodes do {/* g ε L-- REF[p] */if (g ε P-- REF[i] && !Scc-- refs) thencontinue;/* p is a potential entry-- node */web-- nodes := .0.;entry-- nodes := {p};repeat {for each node e in entry-- nodes do{  Expand-- web (&web-- nodes, e, g);}entry-- nodes := .0.;web-- is-- legal :=Check-- if-- web-- is-- legal(web-- nodes,&entry-- nodes);} until web-- is-- legal;/* see if this web can be merged with  previously found webs */Merge-- webs (web-- nodes);}/*Now consider L-- REF nodes in cycles. Note that noneof the nodes in cycles would have been identifiedas entry nodes since (g ε P-- REF[i]) for all nodes ina cycle.*/if (!Scc-- refs && Ref-- nodes ≠ .0.) then {Scc-- refs := true;goto analyze-- refs;}Add-- to-- web-- table(g); }Expand-- web (web-- nodes, p, g){/* This routine recursively finds all successor nodes  of p that should belong to the web for global g.  These nodes are passed back through the web-- nodes   reference parameter.*/web-- nodes += { p };if (p ε Ref-- nodes) then {  Ref-- nodes := Ref-- nodes - {p};}for each successor s of p do {  if (((g ε C-- REF[s]) ∥ (g ε L--REF[s])) &&(s   web-- nodes)) then {Expand-- webs (web-- nodes, s, g);}}}BOOLEAN Check-- if-- web-- is-- legal(web--nodes, new-- entry-- nodes);{/*  This routine checks if any web-- node has both  INTERNAL and EXTERNAL predecessors. It returns  false if this is the case. All external predecessors  of such web-- nodes should be made part of the web and  are passed back as new new-- entry-- nodes.*/legal-- web := true;for each node i in web-- nodes do {Int-- pred := Ext-- pred := false;Ext-- pred-- nodes := .0.;for each predecessor p of node i do {if ( p ε web-- nodes ) then {  Int-- pred = true;}else {  Ext-- pred = true;  Ext-- pred-- nodes = Ext-- pred-- nodes ∪ {p };}}if (Int-- pred && Ext-- pred) then {legal-- web := false;new-- entry-- nodes := new-- entry-- nodes ∪   Ext-- pred-- nodes;}}}Merge-- webs (new-- web){/*  This routine maintains a temporary list of webs for the  global variable currently being analyzed. It checks to  see if the newly identified web has any nodes in common  with the previously identified webs. If so, the webs  are merged as needed such that the new list of webs are  disjoint. Note that merging 2 webs that are each legal  results in a merged web that continues to be legal.*/for each web w of g in Temp-- Web-- List do {if ( (new-- web ∩ w) ≠ .0.) then {   new-- web := w ∪ new-- web;   delete w from Temp-- Web-- List;}}Add new-- web to Temp-- Web-- List;}Add-- to-- web-- table(g){/*  This routine transfers the webs on the Temp-- Web-- List  to the Web.sub. -- Table[ ], discarding webs that would not  be worthwhile to to promote to a register. It also  figures out the true entry-- nodes for each web.*/for each web w of g in Temp-- Web-- List do {/* use some heuristics to decide if this web is worthy  of promotion*/if (web is unworthy of promotion) then  continue;entry-- nodes := .0.;for each node i in w do  Add each predecessor p of node i not in w   to entry-- nodes;web-- id := next available entry in Web-- Table[ ];Web-- Table[web-- id].nodes := w;Web-- Table[web-- id].Entry-- nodes := entry--nodes;Web-- Table[web-- id].intf := .0.;Web-- Table[web-- id].reg := 0;}}The remaining pseudo-code routines are not specified indetail because they use generally available graph coloringalgorithms.Compute-- Web-- Interferences ( ){/*Computes the web interference graph.*/For each entry w1 in the Web-- Table[ ]For each entry w2 in the Web-- Table[ ]  If ((Web-- Table[w1].nodes ∩ Web-- Table[w2].nodes)≠ .0.)  {/* Add a mutual interference between w1 & w2  in the Web-- Interferences[ ] Table.*/Web-- Table[w1].intf ∪= { w2 };Web-- Table[w2].intf ∪= { w1 };}}Color-- Webs ( ){For each web of every global g, compute the web savingsvalue if that web were to be promoted to a register, usingsome heuristics. For example, webs with many entry nodesand few internal nodes that reference the global shouldbe assigned a low savings. Use profile information toassist in identifying heavily accessed nodes and/orvariables if it is available and selected by the user;Sort the webs for every global g, based on the web savingsvalue;/*Color the Web Interference Graph using a fixednumber of interprocedural callee-saves registers byvisiting the Webs in the Sort order determined inStep 1.If a web can't be colored, skip over it.*/for each web w in sort order do {available-- regs := fixed subset of callee-- saves regs;for each web x in Web-- Table[w].intf do {  if (Web-- Table[x].reg ≠ 0) then   available-- regs - = Web-- Table[x].reg;}if (available-- reg ≠ .0.) then  Web-- Table[w].reg := first available-- reg;}}Compute-- Reg-- Actions ( ){/*  This routine computes register actions described  below for the nodes of the call graph. These  register actions are written to the program database,  read by the compiler phase 2 and applied to each  procedure by the register allocator.  Register Action 1:Callee-- saves register X is not available for intraprocedural use (i.e. it is an interprocedural register)Register Action 2 :Insert code to preserve the original contents of interprocedural register XInsert code to load up global variable g into interprocedural register X at the entry point.Insert code to store the global variable g from interprocedural X at the exit point. (i.e. this is a register action to be applied at web entry node, e.g. for variable g1 and register r17 at procedure B in FIG. 5) Note that the load/store instruction can be generated from the information in the program database.Register Action 3 :Convert all memory references to global variable Y into register references to the interprocedural register X (e.g. variable g1 and register r17 in FIG. 5).*/for each web w of each global variable g do {if (Web-- Table[w].reg ≠ 0) then {for each node i in Web-- Table[w].nodes do {  R-- Action [i,g] := { 1 };  if (i ε Web-- Table[2].Entry-- nodes) then   R-- Action [i,g]:=    R-- Action[i,g] ∪ { 2 };  if ( g ε L-- REF[i] ) then   R-- Action [i,g] :=    R-- action[i,g] ∪ { 3 };}}}}______________________________________

Register Spill Optimization

Software procedure calling conventions designate two classes of registers, callee-saves and caller-saves registers. Callee-saves registers that are used must be spilled to memory at procedure entry points and restored at exit points. These registers may be used to keep live values across calls to other procedures. Caller-saves registers cannot hold live values across calls, so their contents must be temporarily saved in memory if those contents are needed after the call.

The simple idea behind register spill optimization is to move callee-saves register spill upwards in the call graph so that descendant nodes may use them "for free", hence the term, spill code motion.

The method according to the invention has the following features:

Regions of the call graph called clusters are identified over which spill code motion may be effective.

The callee-saves registers are pre-allocated to nodes of the cluster by the program analyzer. This allows the register allocator to have some knowledge of every procedure's register usage without forcing the procedures to be compiled in any particular order.

Calls may still be made to procedures that are not known until link time or load time, as long as those procedures follow the standard register usage conventions.

In some procedures, a larger number of caller-saves registers will be available than is allowed by the standard convention.

The available set of callee-saves registers should get utilized more efficiently.

Clusters

Clusters are identified for two reasons. First, to identify the nodes where it is safe and correct to execute spill code for other procedures. Second, to execute spill code relatively infrequently in order to reduce overhead and achieve a performance improvement.

Conceptually, a cluster is a collection of nodes in the call graph that can be viewed as a single entity with regards to register allocation. The procedure calling convention will be adhered to at the boundary of a cluster, but not internally. The idea is to have some nodes within the cluster be able to use callee-saves registers without incurring the expense of saving and restoring them on entry and exit respectively. If such nodes within the cluster are heavily called, then a measurable performance improvement should result.

Breaking up an entire call graph into disjoint clusters of nodes enables more effective pre-allocation of callee-saves registers within smaller regions. Without the notion of clusters, one would run out of the few callee-saves registers quickly in large programs. Regardless of which way callee-saves registers are pre-allocated (top-down or bottom-up), certain regions of the call graph would be deprived of any benefit (or possibly even impacted negatively).

Ideally clusters should be shallow with many nodes close to the cluster root node. In addition, the root node should be invoked less frequently than the internal nodes. Propagating up the callee-spill code within the cluster to the root node should then speed up the application. One is less likely to run out of callee-saves registers for clusters that are short. Additionally, wide clusters will allow the allocation of the same set of callee-saves registers to many sibling nodes within the cluster.

Finally, partitioning the call graph into many different small clusters is preferable to having a few large clusters. This should allow a more uniform distribution of entry-spill code across the entire call graph. Also, the pre-allocation algorithm can try to move entry-spill code across clusters if desired.

Cluster Definition

A cluster in the call graph is defined as a set of nodes with the following properties:

[1] There exists some node R (e.g. A in FIG. 7), called the root of the cluster, which dominates all other nodes within the cluster. (Node D dominates node N if and only if every path from any start node to N includes D.) Note that this does not imply that all nodes dominated by the root node are in the cluster.

[2] For every node P (e.g. E in FIG. 7) in the cluster except R, all immediate predecessors of P (e.g. B in FIG. 7) are also in the cluster.

[3] A non-root node P is included only in the cluster of the immediately dominating root node.

For spill code motion to improve performance, the root node of a cluster should be called less frequently than the internal nodes of the cluster. This can be estimated by associating weights on the edges of the call graph which indicate relative call frequencies. These weights can be heuristically derived by the compiler first phase, but they may be more accurately assigned by using profile information.

Root nodes are selected by examining the incoming and outgoing calls at each node. The successors of a root node are added to the cluster by applying conditions [1] and [2]. Note that condition [3] allows a leaf node of one cluster to also be the root node of another cluster (e.g. C and D in FIG. 7).

Register Pre-Allocation

The register need analyzer of the compiler first phase can communicate the approximate callee-saves register requirements to the Program Analyzer by performing normal intraprocedural register allocation on the intermediate code generated. Based on this information for each node of a cluster, registers are allocated within each cluster starting at the root of the cluster and working downwards.

In effect, the callee-saves registers are split into four classes which describe how each register can be used within each procedure. The classes are defined by two conditions: 1) whether or not the register must be spilled on entry and restored on exit when it is used in the procedure and 2) whether or not the register can be used to hold values across procedure calls. These classes are identified by the following sets:

FREE[P]--registers in this set need not be spilled if they are used, and may hold live values across calls. These are essentially the interprocedural registers.

CALLER-- SAVES[P]--registers in this set need not be spilled if they are used, but may not hold live values across calls.

CALLEE-- SAVES[P]--registers in this set must be spilled if they are used, and may hold live values across calls.

MSPILL[P]--registers in this set must be spilled if they are used and they may not hold live values across calls.

When register assignment is done by the register allocator 62 (FIG. 3) for a procedure, the register allocator will query the program database 56 to identify which set each machine register belongs to, and replace each pseudo-register 102, 104, 106 with a machine register 112, 113, 114 from the appropriate set. The register allocator must also add spill code 116, 118 for those machine registers assigned from the CALLEE-- SAVES and MSPILL sets.

There is an additional requirement in the register allocator that all registers in in the MSPILL set at a cluster root procedure must be spilled on entry and restored on exit (spill code 116 and 118 in FIG. 3), regardless of whether or not they are actually used inside that procedure. This will accomplish the goal of having the root node execute the spill code for the remaining nodes of the cluster. The algorithms described hereinafter arrange that MSPILL will always be empty at non-root nodes belonging to a cluster.

The algorithms described hereinafter also eliminate the possibility of recursive call cycles from occurring within a cluster. Consider the simple case of a self-recursive routine which uses callee-saves registers. Such a routine would expect that the values in these registers would remain safe across the recursive call. If spill code is not executed at the entry point, however, the values that were live across that recursive call will be destroyed.

Recursive cycles are prevented from occurring within clusters. However, entire clusters may occur within cycles in the call graph.

If there exists a node within a cluster that makes a call back to the root node, pre-allocation will note be done on that cluster. This is done for performance reasons.

The pre-allocation technique herein proposed has one major vulnerability. In particular, if a selected cluster root node is a procedure that is called more frequently than the other nodes of the cluster, calls to that procedure would be slowed down.

FIG. 7 shows three clusters 158, 160, 162 that might be identified for the program call graph 130. Cluster 158 comprises the nodes A, B, C, D, and E. Cluster 160 comprises the nodes D and E. Cluster 162 comprises the nodes C, F, G, and I. Note that nodes D and C are cluster root nodes that are themselves part of cluster 158. Node A is the root node for cluster 158.

Spill Optimization Algorithms

Input: A call graph G of `n` nodes. It is assumed that the representation of the call graph allows easy identification of a node's successors & predecessors.

Output: Register sets MSPILL, CALLEE-- SAVES, CALLER-- SAVES, and FREE for each node of G to be used by the register allocator. This information is stored in the program database 56.

The following data structures are used to help identify clusters.

Data Structures

visited [1. . . n] Flag to mark a node as having been examined.

dom [1. . . n] For each node i, dom[i] represents the set of nodes that dominate it.

cluster [1. . . n] For each ROOT node i, cluster[i] lists the set of associated cluster nodes.

AVAIL [1. . . n] For each node i, AVAIL[i] represents a set of registers that will be used in the register preallocation process.

______________________________________Pseudo-- code:register-- preallocation( G ) {/* G is a call graph */Identify-- Clusters( );Initialize the MSPILL and FREE sets at all nodes in Gto be empty;Initialize the CALLEE-- SAVES and CALLER-- SAVES sets ateach node to the standard procedure call register sets;Initialize visited[*] := false;for each cluster root node P in G in depth first order do/* don't allow recursive calls back to the root */if any of immediate predecessor of P is an element  of the cluster rooted at P then  continue;Make a pre-pass over the cluster to see whichregisters are used for global variables and todetermine which registers are used by clustersbelow this one;AVAILABLE-- REGS := CALLEE-- SAVES[P]  (global interprocedural regs used in cluster);Assign priorities to the registers based on whichregisters are used by clusters whose roots aremembers of this cluster. Registers not used by thesechild clusters have highest priority;/* USED is a global set of registers */USED := Get-- Regs-- From-- Set( P, AVAILABLE-- REGS);/* * USED is now the set of callee-saves registers * needed by the root node itself. */AVAIL[P] := AVAILABLE-- REGS - USED;CALLEE-- SAVES[P] := USED;Prealloc-- Node ( P );MSPILL[P] := USED;if P is a member of another cluster then  visited[P] := false;for each node Q in the cluster do  if Q is not a cluster root then{/*Move into CALLER-- SAVES any segments that arecurrently in AVAIL, as long as they arealso in MSPILL at the root node of thecluster.*/CALLER-- SAVES[Q] := AVAIL[Q] ∩ MSPILL[P];MSPILL[Q] := .0. ;}}}Identify-- Clusters ( ){compute dom[ i] set for all nodes of G;Calc-- clusters ( );}Calc-- clusters ( ){set visited[i] to false at all nodes of G;set cluster[i] to .0. at all nodes of G;for each start node S do {  Examine-- node( S , NULL );}}Examine-- node( P, CurClusterRoot ){/* * EXTERNAL-- NODE(P) is true for any node P which * is not included in the source code of the application. */if (EXTERNAL-- NODE(P) ∥ postpone-- visit(P))return;visited(P) := true;if (   CurClusterRoot != NULL &&  All predecessors of P have been visited &&  CurClusterRoot dominates P &&  All predecessors of P are in the cluster  rooted at CurClusterRootthen {add node P to the cluster rooted at CurClusterRoot ;}if (Is-- A-- Root( P )) then { Create a cluster rooted at P; CurClusterRoot := P;}for each successor S of P do {if (!visited(S))  Examine-- node( S, CurClusterRoot );}}BOOLEAN postpone-- visit( S ){/*This function will return false if eithera)     all predecessors of S have been visitedOR -b)     Incoming edges from all predecessors nodes  that haven't been visited are BACK edges.Conversely, this function will return true ifa)     Not all predecessors of S have been visitedAND -b)     there exists at least one unvisited predecessor  P, and the edge from P to S is not a BACK edge  in the graph (i.e. it's a FORWARD edge, CROSS  edge, or TREE edge).This function is defined for two reasons. First,visiting a node should be delayed if not all of it'spredecessors have been visited yet. Second, even ifnot all predecessors have been visited, there are caseswhere one must select some node to visit (or else therewill be a deadlock situation.) This happens when theincoming edges to S are "back edges" (edges fromdescendents of S.)See the discussion below for an example of why this isnecessary.*/}Is-- A-- Root (i){Return true if should be a root node, false otherwise.This is a heuristic based on the incoming andoutgoing calls at node i. If available and selected bythe user, use profile information to get accurate countsof incoming and outgoing call counts at node i.}Get-- Regs-- From-- Set( P, REGS ){Let N be the number of callee-saves registersrequested for procedure P.This function selects up to N registers from the setREGS, using the priority determined in the prepass todetermine in what order the elements should be selected.The function returns the set of registers that wereselected. Note that the result set could be empty.}Prealloc-- Node( P ){visited[P] := true;TEMP := the intersection of the AVAIL sets over all    immediate predecessors of P./* TEMP is now the set of register segments that  are guaranteed not to hold live values on entry to P.*/if (P is the root of a cluster) then{/*First, try to move any registers that are inMSPILL[P] into the MSPILL of the current clusterroot. Then, try to move any registers that are inCALLEE-- SAVES[P] into the FREE[P] if possible.*/USED := USED ∪ ( MSPILL[P] ∩ TEMP ) ;TEMP ;[P] := NSPILL[P]USED := USED ∪ ( CALLEE-- SAVES[P] ∩ TEMP ) ;FREE[P] := CALLEE-- SAVES[P] ∩ TEMP ;FREE[P] ;b.-- SAVES[P] := CALLEE-- SAVES[P]}else{FREE[P] := Get-- Regs-- From-- Set( P, TEMP ) ;AVAIL[P] := TEMP - FREE[P] ;CALLEE-- SAVES[P] := CALLEE-- SAVES[P](FREE[P] ∪ AVAIL[P]) ;USED := USED ∪ FREE[P] ;};for each immediate successor S of P do{if S is in the current cluster andall immediate predecessors of S are visited thenPrealloc-- Node( S );}};______________________________________
An Example

Consider the sample call graph 170 shown in FIG. 8. The nodes of this sample call graph are interconnected by directional edges 180, 182, 184, 186, 188, 190. Assume that the depth-first search order of the nodes of this call graph is J, K, L, M, N, P. The depth-first search order is the order in which these nodes are first considered by the algorithm.

Suppose that the function Is-- A-- Root() returns true for nodes J and M. After visiting node J, Examine-- node() will be called on node K. At this point, however, the decision on whether or not to add node K to the cluster rooted at node J will be postponed. This is because node P has not yet been visited.

If node K is added to the cluster rooted at J, and P is added to the cluster roted at M, condition [2] from the cluster definition will be violated. (Node K will have an immediate predecessor not in its cluster.)

The invention is practiced as follows: Summary files 48, 50, 52 are produced for each of the source files 18, 20, 22. The program analyzer 54 is invoked specifying the names of the summary files as command line options. The program analyzer 54 reads each of the summary files, constructs the program call graph 130, computes register actions needed to do global variable promotion and the register sets (MSPILL, FREE, CALLER-- SAVES, CALLEE-- SAVES) for spill code motion and finally writes these register actions and register sets out to the program database 56.

The program analyzer 54 performs global register promotion and register spill optimization independently. Interprocedural registers reserved for global promotion are not available for pre-allocation.

After the program database 56 is created, the intermediate files (or the source files) are read by the compiler second phase 12B along with the record of register actions and register sets from the program database for each procedure being compiled. The register actions are then applied to each procedure by the register allocator 62 and object files 120 are created. The object files 120 thus created are read by the program linker 17 linked with the run-time libraries 19 to produce an executable file 24 with interprocedural register allocation.

File Structure

The logical structure of the summary files is described below using Backus Normal Format (BNF) notation.

______________________________________<Summary-- File> ::= <Procedure-- Record> *<Procedure-- Record> ::=  Procedure Name     +  Num-- Callee-- Regs-- Needed                     +  <Callee-- Info-- Record>*                     +  <Global-- Usage-- Info-- Record>*                     +  <Plabel-- Record>*<Callee-- Info-- Record> ::=  Callee-- Name +  Static-- Call-- Count<Global-- Usage-- Info-- Record>::=  Global-- Variable-- Name                     +  Global-- Attributes                     +  Weighted-- Reference-- Count                     +  Num-- References                     +<Global-- Attributes> ::=  Global-- Type |  Static-- Variable?                     |  Address-- Taken?<Plabel-- Record> ::= Indirectly-- callable-- procedure.sub.-- name______________________________________

The variable usage analyzer 40 computes the Global-- Usage-- Info-- Record. The procedure call analyzer 44 computes the Callee-- Info-- Record and Plabel-- Record. The register need analyzer computes the Num-- Callee-- Regs-- Needed field of the Procedure-- Record. Finally, the summary file generator 46 organizes the all the fields of the Procedure-- Record and writes it out to the summary file 48, 50, 52 of the corresponding source code file 18, 20, 22.

The logical structure of the program database file is described below using the same BNF notation.

______________________________________<Database-- File>  ::=          <Procedure-- Record>*<Procedure-- Record> ::=          <Register-- Set>                            +          <Reg-- Action-- Record><Register-- Set> ::=          MSPILL            +          FREE              +          CALLER-- SAVES                            +          CALLEE-- SAVES<Reg-- Action-- Record> ::=            Reg-- Action [1 . . . 3]                            +            Global-- Name                            +            Interprocedural Register______________________________________

The invention has now been described with respect to specific embodiments. Other embodiments will be apparent to those of ordinary skill in this art upon reference to this description. It is therefore not intended that this invention be limited, except as limited by the appended claims.

Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US4571678 *Nov 5, 1982Feb 18, 1986International Business Machines CorporationRegister allocation and spilling via graph coloring
US4642764 *Aug 13, 1984Feb 10, 1987International Business Machines CorporationMethod of developing formal identities and program bases in an optimizing compiler
US4656582 *Feb 4, 1985Apr 7, 1987International Business Machines CorporationGenerating storage reference instructions in an optimizing compiler
US4777588 *Aug 30, 1985Oct 11, 1988Advanced Micro Devices, Inc.General-purpose register file optimized for intraprocedural register allocation, procedure calls, and multitasking performance
US4782444 *Dec 17, 1985Nov 1, 1988International Business Machine CorporationCompilation using two-colored pebbling register allocation method such that spill code amount is invariant with basic block's textual ordering
US4953084 *Nov 16, 1987Aug 28, 1990Hewlett-Packard CompanyMethod and apparatus using variable ranges to support symbolic debugging of optimized code
US4953106 *May 23, 1989Aug 28, 1990At&T Bell LaboratoriesTechnique for drawing directed graphs
US4961141 *Dec 16, 1988Oct 2, 1990International Business Machines CorporationGenerating efficient code for a computer with dissimilar register spaces
US5083263 *Jul 28, 1988Jan 21, 1992Sun Microsystems, Inc.BISC with interconnected register ring and selectively operating portion of the ring as a conventional computer
US5161216 *Mar 8, 1989Nov 3, 1992Wisconsin Alumni Research FoundationInterprocedural slicing of computer programs using dependence graphs
Non-Patent Citations
Reference
1Fred C. Chow, "Minimizing Register Usage Penalty at Procedure Calls", Proceedings of the SIGPLAN, '88 Conference on Programming Language Design and Implementation, Jul. 1988, pp. 85-94.
2 *Fred C. Chow, Minimizing Register Usage Penalty at Procedure Calls , Proceedings of the SIGPLAN , 88 Conference on Programming Language Design and Implementation, Jul. 1988, pp. 85 94.
3Hans Mulder, "Data Buffering: Run-Time Versus Compile Time Support", Proceedings of the 3rd International Conference on Architectural Support for Programming Languages and Operating Systems, Apr. 3-6, 1989, pp. 144-151.
4 *Hans Mulder, Data Buffering: Run Time Versus Compile Time Support , Proceedings of the 3rd International Conference on Architectural Support for Programming Languages and Operating Systems , Apr. 3 6, 1989, pp. 144 151.
5Keith D. Cooper, et al., "The Impact of Interprocedural Analysis and Optimization in Rn Programming Environment", Published in ACM Transactions on Programming Languages and Systems, Oct. 1986, pp. 491-523.
6 *Keith D. Cooper, et al., The Impact of Interprocedural Analysis and Optimization in R n Programming Environment , Published in ACM Transactions on Programming Languages and Systems, Oct. 1986, pp. 491 523.
7Mark I. Himelstein, et al., "Cross-Module Optimizations: Its Implementation and Benefits", Proceedings of the Summer 1987 USENIX Conference, pp. 347-356.
8 *Mark I. Himelstein, et al., Cross Module Optimizations: Its Implementation and Benefits , Proceedings of the Summer 1987 USENIX Conference , pp. 347 356.
9P. A. Steenkiste, "LISP on a Reduced Instruction Set Processor: Characterization and Optimization", Stanford University Computer Systems Laboratory, PhD Thesis, Chapter 5, Mar. 1987.
10 *P. A. Steenkiste, LISP on a Reduced Instruction Set Processor: Characterization and Optimization , Stanford University Computer Systems Laboratory, PhD Thesis, Chapter 5, Mar. 1987.
11Wall, "Global Register Allocation at Link Time", Proc. of the SIGPLAN, SIGPLAN Notices, vol. 21, No. 7, Jul. 1986, pp. 264-275.
12 *Wall, Global Register Allocation at Link Time , Proc. of the SIGPLAN, SIGPLAN Notices, vol. 21, No. 7, Jul. 1986, pp. 264 275.
Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US5530866 *Oct 13, 1994Jun 25, 1996Tera Computer CompanyRegister allocation methods having upward pass for determining and propagating variable usage information and downward pass for binding; both passes utilizing interference graphs via coloring
US5535394 *Oct 10, 1995Jul 9, 1996International Business Machines CorporationMethods for distinguishing dynamically allocated storage along different call paths and for providing a compact representation of alias analysis
US5551040 *Jan 9, 1995Aug 27, 1996Lucent Technologies Inc.Methods and apparatus for specifying the contexts of names in callback-style programming
US5555417 *Jan 23, 1995Sep 10, 1996Hewlett-Packard CompanyMethod and apparatus for compiling computer programs with interprocedural register allocation
US5671419 *Jun 15, 1995Sep 23, 1997International Business Machines CorporationInterprocedural data-flow analysis that supports recursion while only performing one flow-sensitive analysis of each procedure
US5768592 *Sep 27, 1994Jun 16, 1998Intel CorporationMethod and apparatus for managing profile data
US5778212 *Jun 3, 1996Jul 7, 1998Silicon Graphics, Inc.Interprocedural analysis user interface
US5778232 *Jul 3, 1996Jul 7, 1998Hewlett-Packard CompanyAutomatic compiler restructuring of COBOL programs into a proc per paragraph model
US5787284 *Oct 7, 1996Jul 28, 1998International Business Machines CorporationImproving memory layout based on connectivity considerations
US5797012 *Oct 7, 1996Aug 18, 1998International Business Machines CorporationConnectivity based program partitioning
US5819064 *Nov 8, 1995Oct 6, 1998President And Fellows Of Harvard CollegeHardware extraction technique for programmable reduced instruction set computers
US5828886 *Feb 23, 1995Oct 27, 1998Fujitsu LimitedCompiling apparatus and method for promoting an optimization effect of a program
US5850549 *Oct 7, 1996Dec 15, 1998International Business Machines CorporationGlobal variable coalescing
US5857103 *Jun 14, 1996Jan 5, 1999Sun Microsystems, Inc.Method and apparatus for addressing extended registers on a processor in a computer system
US5907711 *Jan 22, 1996May 25, 1999Hewlett-Packard CompanyMethod and apparatus for transforming multiplications into product table lookup references
US5920723 *Feb 5, 1997Jul 6, 1999Hewlett-Packard CompanyCompiler with inter-modular procedure optimization
US5950009 *Mar 10, 1997Sep 7, 1999International Business Machines CoporationMethod and apparatus for profile-based reordering of program portions in a computer program
US6026242 *May 29, 1998Feb 15, 2000Sun Microsystems, Inc.Methods, computer program products, and apparatus for initializing global registers
US6029004 *Mar 17, 1997Feb 22, 2000International Business Machines CorporationMethod and apparatus for modular reordering of portions of a computer program based on profile data
US6139200 *Jun 30, 1997Oct 31, 2000Sun Microsystems, Inc.Register resource allocation feedback
US6175957 *Dec 9, 1997Jan 16, 2001International Business Machines CorporationMethod of, system for, and computer program product for providing efficient utilization of memory hierarchy through code restructuring
US6226787Jan 25, 1999May 1, 2001Hewlett-Packard CompanyVisualization method and system for dynamically displaying operations of a program
US6295640 *May 8, 1998Sep 25, 2001Apple Computer, Inc.Method and apparatus for distinguishing reference values from non-reference values in a runtime environment
US6434743 *Apr 23, 1999Aug 13, 2002Sun Microsystems, Inc.Method and apparatus for allocating stack slots
US6453466 *Aug 13, 2001Sep 17, 2002Apple Computer, Inc.Method and apparatus for distinguishing reference values from non-reference values in a runtime environment
US6631518 *Mar 19, 1997Oct 7, 2003International Business Machines CorporationGenerating and utilizing organized profile information
US6684392 *Jun 27, 2002Jan 27, 2004Apple Computer, Inc.Method and apparatus for distinguishing reference values from non-reference values in a runtime environment
US6738966 *Jul 7, 2000May 18, 2004Matsushita Electric Industrial Co., Ltd.Compiling device, computer-readable recording medium on which a compiling program is recorded and a compiling method
US6839895Oct 10, 2000Jan 4, 2005International Business Machines CorporationMethod of, system for, and computer program product for providing efficient utilization of memory hierarchy through code restructuring
US6848098 *Jul 26, 2000Jan 25, 2005Fujitsu LimitedProcess and apparatus for optimizing program written in interpretive language for each callable program part base on frequency of executions
US7003507 *Mar 30, 2001Feb 21, 2006International Business Machines CorporationScalable propagation-based methods for call graph construction
US7069548 *Jun 28, 2002Jun 27, 2006Intel CorporationInter-procedure global register allocation method
US7100156 *Sep 27, 2001Aug 29, 2006International Business Machines CorporationInterprocedural dead store elimination
US7207032Mar 28, 2003Apr 17, 2007Applied Micro Circuits CorporationExpanding a software program by insertion of statements
US7219334 *Feb 26, 2003May 15, 2007International Business Machines CorporationProgram conversion method, data processing apparatus and program
US7269827 *Oct 21, 2002Sep 11, 2007Intel CorporationMethod and apparatus for compiling code
US7278137Dec 26, 2002Oct 2, 2007Arc InternationalMethods and apparatus for compiling instructions for a data processor
US7302679Oct 31, 2003Nov 27, 2007Hewlett-Packard Development Company, L.P.Scalable cross-file inlining through locality-based transformation ordering
US7305665 *Jun 11, 2003Dec 4, 2007International Business Machines CorporationCompiler register allocation and compilation
US7543020Feb 17, 2005Jun 2, 2009Cisco Technology, Inc.Distributed client services based on execution of service attributes and data attributes by multiple nodes in resource groups
US7590977 *Oct 13, 2005Sep 15, 2009International Business Machines CorporationMethod and system for reducing memory reference overhead associated with threadprivate variables in parallel programs
US7620946 *Nov 17, 2009Jeffry Thomas RussellProgram slicing for codesign of embedded systems
US7739484 *Oct 5, 2007Jun 15, 2010Mips Technologies, Inc.Instruction encoding to indicate whether to store argument registers as static registers and return address in subroutine stack
US7818731 *Oct 19, 2010International Business Machines CorporationMethod and system for reducing memory reference overhead associated with treadprivate variables in parallel programs
US7991835Aug 2, 2011Cisco Technology, Inc.Distributed client services based on execution of service attributes and data attributes by multiple nodes in resource groups
US7996825Aug 9, 2011Hewlett-Packard Development Company, L.P.Cross-file inlining by using summaries and global worklist
US8051170Nov 1, 2011Cisco Technology, Inc.Distributed computing based on multiple nodes with determined capacity selectively joining resource groups having resource requirements
US8104026Oct 29, 2007Jan 24, 2012International Business Machines CorporationCompiler register allocation and compilation
US8156471 *Nov 9, 2001Apr 10, 2012Oracle International CorporationMulti-language execution method
US8166450Oct 1, 2007Apr 24, 2012Synopsys, Inc.Methods and apparatus for compiling instructions for a data processor
US8195742Jun 5, 2012Cisco Technology, Inc.Distributed client services based on execution of service attributes and data attributes by multiple nodes in resource groups
US8239540Aug 7, 2012Cisco Technology, Inc.Distributed computing based on multiple nodes with determined capacity selectively joining resource groups having resource requirements
US8312429Dec 19, 2000Nov 13, 2012Oracle International CorporationCell based data processing
US8468511 *Dec 31, 2008Jun 18, 2013International Business Machines CorporationUse of name mangling techniques to encode cross procedure register assignment
US8555267 *Mar 3, 2010Oct 8, 2013Red Hat, Inc.Performing register allocation of program variables based on priority spills and assignments
US8639816Jul 3, 2012Jan 28, 2014Cisco Technology, Inc.Distributed computing based on multiple nodes with determined capacity selectively joining resource groups having resource requirements
US8671399 *Oct 31, 2007Mar 11, 2014International Business Machines CorporationSystems and methods for efficiently using stack registers and storage medium therefor
US9015687 *Mar 30, 2011Apr 21, 2015Intel CorporationRegister liveness analysis for SIMD architectures
US20020095669 *Sep 27, 2001Jul 18, 2002Archambault Roch GeorgesInterprocedural dead store elimination
US20020133808 *Dec 19, 2000Sep 19, 2002Adam BosworthCell based data processing
US20030018603 *Mar 30, 2001Jan 23, 2003International Business Machines CorporationScalable propagation-based methods for call graph construction
US20030177479 *Feb 26, 2003Sep 18, 2003International Business Machines CorporationProgram conversion method, data processing apparatus and program
US20040003385 *Jun 28, 2002Jan 1, 2004Intel CorporationInter-procedure global register allocation method
US20040010784 *Jun 11, 2003Jan 15, 2004International Business Machines CorporationCompiler register allocation and compilation
US20040040011 *Nov 9, 2001Feb 26, 2004Adam BosworthMulti-language execution method
US20040078791 *Oct 21, 2002Apr 22, 2004Metzger Markus T.Method and apparatus for compiling code
US20040226006 *May 5, 2003Nov 11, 2004Jeffry RussellProgram slicing for codesign of embedded systems
US20050097527 *Oct 31, 2003May 5, 2005Chakrabarti Dhruva R.Scalable cross-file inlining through locality-based transformation ordering
US20050097528 *Oct 31, 2003May 5, 2005Chakrabarti Dhruva R.Cross-file inlining by using summaries and global worklist
US20050097533 *Oct 31, 2003May 5, 2005Chakrabarti Dhruva R.Run-time performance with call site inline specialization
US20050125783 *Dec 9, 2003Jun 9, 2005Texas Instruments IncorporatedProgram optimization with intermediate code
US20060179106 *Feb 10, 2005Aug 10, 2006Turner Bryan CDistributed computing based on multiple nodes with determined capacity selectively joining resource groups having resource requirements
US20060179143 *Feb 17, 2005Aug 10, 2006Walker Douglas JDistributed client services based on execution of service attributes and data attributes by multiple nodes in resource groups
US20070089105 *Oct 13, 2005Apr 19, 2007Archambault Roch GMethod and system for reducing memory reference overhead associated with threadprivate variables in parallel programs
US20080005722 *Jun 28, 2006Jan 3, 2008Hidenori MatsuzakiCompiling device, compiling method and recording medium
US20080028195 *Oct 5, 2007Jan 31, 2008Mips Technologies, Inc.Method and Apparatus for Saving and Restoring Processor Register Values and Allocating and Deallocating Stack Memory
US20080059955 *Oct 31, 2007Mar 6, 2008International Business Machines CorporationCompiling method and storage medium therefor
US20080229297 *May 29, 2008Sep 18, 2008International Business Machines CorporationMethod and system for reducing memory reference overhead associated with treadprivate variables in parallel programs
US20080320246 *Oct 1, 2007Dec 25, 2008Fuhler Richard AMethods and apparatus for compiling instructions for a data processor
US20090193400 *Jul 30, 2009Ivan BaevInterprocedural register allocation for global variables
US20090287825 *Nov 19, 2009Cisco Technology, Inc.Distributed client services based on execution of service attributes and data attributes by multiple nodes in resource groups
US20100169872 *Dec 31, 2008Jul 1, 2010International Business Machines CorporationUse of Name Mangling Techniques To Encode Cross Procedure Register Assignment
US20110167415 *Jul 7, 2011Kabushiki Kaisha ToshibaLanguage processing apparatus, language processing method, and computer program product
US20110219364 *Mar 3, 2010Sep 8, 2011Vladimir MakarovMechanism for Performing Register Allocation of Program Variables Based on Priority Spills and Assignments
US20120254847 *Mar 30, 2011Oct 4, 2012Biju GeorgeRegister liveness analysis for simd architectures
Classifications
U.S. Classification717/157
International ClassificationG06F9/45
Cooperative ClassificationG06F8/441
European ClassificationG06F8/441
Legal Events
DateCodeEventDescription
Dec 7, 1989ASAssignment
Owner name: HEWLETT-PACKARD COMPANY, A CORP. OF CA, CALIFORNI
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST.;ASSIGNORS:ODNERT, DARYL;SANTHANAM, VATSA;REEL/FRAME:005196/0244
Effective date: 19891113
Dec 24, 1998FPAYFee payment
Year of fee payment: 4
Jan 19, 1999REMIMaintenance fee reminder mailed
Jan 16, 2001ASAssignment
Owner name: HEWLETT-PACKARD COMPANY, COLORADO
Free format text: MERGER;ASSIGNOR:HEWLETT-PACKARD COMPANY;REEL/FRAME:011523/0469
Effective date: 19980520
Sep 23, 2002FPAYFee payment
Year of fee payment: 8
Dec 27, 2006FPAYFee payment
Year of fee payment: 12